Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 ·...
Transcript of Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 ·...
![Page 1: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/1.jpg)
Local Search
Local Search (Greedy Descent):
Maintain an assignment of a value to each variable.
Repeat:
I Select a variable to changeI Select a new value for that variable
Until a satisfying assignment is found
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 1
![Page 2: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/2.jpg)
Local Search for CSPs
Aim: find an assignment with zero unsatisfied constraints.
Given an assignment of a value to each variable, a conflict isan unsatisfied constraint.
The goal is an assignment with zero conflicts.
Heuristic function to be minimized: the number of conflicts.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 2
![Page 3: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/3.jpg)
Greedy Descent Variants
To choose a variable to change and a new value for it:
Find a variable-value pair that minimizes the number ofconflicts
Select a variable that participates in the most conflicts.Select a value that minimizes the number of conflicts.
Select a variable that appears in any conflict.Select a value that minimizes the number of conflicts.
Select a variable at random.Select a value that minimizes the number of conflicts.
Select a variable and value at random; accept this change if itdoesn’t increase the number of conflicts.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 3
![Page 4: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/4.jpg)
Complex Domains
When the domains are small or unordered, the neighbors of anassignment can correspond to choosing another value for oneof the variables.
When the domains are large and ordered, the neighbors of anassignment are the adjacent values for one of the variables.
If the domains are continuous, Gradient descent changeseach variable proportional to the gradient of the heuristicfunction in that direction.The value of variable Xi goes from vi to
vi − η ∂h∂Xi
.η is the step size.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 4
![Page 5: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/5.jpg)
Complex Domains
When the domains are small or unordered, the neighbors of anassignment can correspond to choosing another value for oneof the variables.
When the domains are large and ordered, the neighbors of anassignment are the adjacent values for one of the variables.
If the domains are continuous, Gradient descent changeseach variable proportional to the gradient of the heuristicfunction in that direction.The value of variable Xi goes from vi to vi − η ∂h
∂Xi.
η is the step size.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 5
![Page 6: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/6.jpg)
Problems with Greedy Descent
a local minimum that isnot a global minimum
a plateau where theheuristic values areuninformative
a ridge is a localminimum where n-steplook-ahead might help
Ridge
Local Minimum
Plateau
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 6
![Page 7: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/7.jpg)
Randomized Algorithms
Consider two methods to find a minimum value:I Greedy descent, starting from some position, keep moving
down & report minimum value foundI Pick values at random & report minimum value found
Which do you expect to work better to find a globalminimum?
Can a mix work better?
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 7
![Page 8: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/8.jpg)
Randomized Greedy Descent
As well as downward steps we can allow for:
Random steps: move to a random neighbor.
Random restart: reassign random values to all variables.
Which is more expensive computationally?
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 8
![Page 9: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/9.jpg)
1-Dimensional Ordered Examples
Two 1-dimensional search spaces; step right or left:
(a) (b)
Which method would most easily find the global minimum?
What happens in hundreds or thousands of dimensions?
What if different parts of the search space have differentstructure?
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 9
![Page 10: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/10.jpg)
Stochastic Local Search
Stochastic local search is a mix of:
Greedy descent: move to a lowest neighbor
Random walk: taking some random steps
Random restart: reassigning values to all variables
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 10
![Page 11: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/11.jpg)
Random Walk
Variants of random walk:
When choosing the best variable-value pair, randomlysometimes choose a random variable-value pair.
When selecting a variable then a value:I Sometimes choose any variable that participates in the most
conflicts.I Sometimes choose any variable that participates in any conflict
(a red node).I Sometimes choose any variable.
Sometimes choose the best value and sometimes choose arandom value.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 11
![Page 12: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/12.jpg)
Comparing Stochastic Algorithms
How can you compare three algorithms whenI one solves the problem 30% of the time very quickly but
doesn’t halt for the other 70% of the casesI one solves 60% of the cases reasonably quickly but doesn’t
solve the restI one solves the problem in 100% of the cases, but slowly?
Summary statistics, such as mean run time, median run time,and mode run time don’t make much sense.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 12
![Page 13: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/13.jpg)
Comparing Stochastic Algorithms
How can you compare three algorithms whenI one solves the problem 30% of the time very quickly but
doesn’t halt for the other 70% of the casesI one solves 60% of the cases reasonably quickly but doesn’t
solve the restI one solves the problem in 100% of the cases, but slowly?
Summary statistics, such as mean run time, median run time,and mode run time don’t make much sense.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 13
![Page 14: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/14.jpg)
Runtime Distribution
Plots runtime (or number of steps) and the proportion (ornumber) of the runs that are solved within that runtime.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 10 100 1000
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 14
![Page 15: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/15.jpg)
Variant: Simulated Annealing
Pick a variable at random and a new value at random.
If it is an improvement, adopt it.
If it isn’t an improvement, adopt it probabilistically dependingon a temperature parameter, T .
I With current assignment n and proposed assignment n′ wemove to n′ with probability e(h(n
′)−h(n))/T
Temperature can be reduced.
Probability of accepting a change:
Temperature 1-worse 2-worse 3-worse
10 0.91 0.81 0.741 0.37 0.14 0.050.25 0.02 0.0003 0.0000060.1 0.00005 2× 10−9 9× 10−14
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 15
![Page 16: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/16.jpg)
Variant: Simulated Annealing
Pick a variable at random and a new value at random.
If it is an improvement, adopt it.
If it isn’t an improvement, adopt it probabilistically dependingon a temperature parameter, T .
I With current assignment n and proposed assignment n′ wemove to n′ with probability e(h(n
′)−h(n))/T
Temperature can be reduced.
Probability of accepting a change:
Temperature 1-worse 2-worse 3-worse
10 0.91 0.81 0.741 0.37 0.14 0.050.25 0.02 0.0003 0.0000060.1 0.00005 2× 10−9 9× 10−14
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 16
![Page 17: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/17.jpg)
Tabu lists
To prevent cycling we can maintain a tabu list of the k lastassignments.
Don’t allow an assignment that is already on the tabu list.
If k = 1, we don’t allow an assignment of to the same valueto the variable chosen.
We can implement it more efficiently than as a list ofcomplete assignments.
It can be expensive if k is large.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 17
![Page 18: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/18.jpg)
Parallel Search
A total assignment is called an individual .
Idea: maintain a population of k individuals instead of one.
At every stage, update each individual in the population.
Whenever an individual is a solution, it can be reported.
Like k restarts, but uses k times the minimum number ofsteps.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 18
![Page 19: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/19.jpg)
Beam Search
Like parallel search, with k individuals, but choose the k bestout of all of the neighbors.
When k = 1, it is greedy descent.
When k =∞, it is breadth-first search.
The value of k lets us limit space and parallelism.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 19
![Page 20: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/20.jpg)
Stochastic Beam Search
Like beam search, but it probabilistically chooses the kindividuals at the next generation.
The probability that a neighbor is chosen is proportional to itsheuristic value.
This maintains diversity amongst the individuals.
The heuristic value reflects the fitness of the individual.
Like asexual reproduction: each individual mutates and thefittest ones survive.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 20
![Page 21: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/21.jpg)
Genetic Algorithms
Like stochastic beam search, but pairs of individuals arecombined to create the offspring:
For each generation:I Randomly choose pairs of individuals where the fittest
individuals are more likely to be chosen.I For each pair, perform a cross-over: form two offspring each
taking different parts of their parents:I Mutate some values.
Stop when a solution is found.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 21
![Page 22: Artificial Intelligence, Lecture 4.2, Page 1poole/aibook/slides/ch04/lect2.pdf · 2013-09-16 · Artificial Intelligence, Lecture 4.2, Page 1 Author ©D. Poole and A. Mackworth 2010](https://reader036.fdocuments.net/reader036/viewer/2022081406/5f1203ae92c6b9366c7c0954/html5/thumbnails/22.jpg)
Crossover
Given two individuals:
X1 = a1,X2 = a2, . . . ,Xm = am
X1 = b1,X2 = b2, . . . ,Xm = bm
Select i at random.
Form two offspring:
X1 = a1, . . . ,Xi = ai ,Xi+1 = bi+1, . . . ,Xm = bm
X1 = b1, . . . ,Xi = bi ,Xi+1 = ai+1, . . . ,Xm = am
The effectiveness depends on the ordering of the variables.
Many variations are possible.
c©D. Poole and A. Mackworth 2010 Artificial Intelligence, Lecture 4.2, Page 22